NVIDIA AI Updates: July 27, 2026

1. Nemotron 3 Ultra Tops Open Models on Agentic RTL Coding

NVIDIA. NVIDIA introduced Nemotron 3 Ultra, an open model paired with the ACE-RTL agent for iterative hardware design through generate-test-reflect loops. The 550B-parameter hybrid Mamba-Attention Mixture-of-Experts model (55B active, 1M-token context) posted a 97.1 percent average pass rate on the Comprehensive Verilog Design Problems benchmark, ahead of GLM 5.2 (92.1 percent) and Kimi K2.6 (95.2 percent), while using roughly 71 percent fewer tokens per iteration than Kimi K2.6. For practitioners, the results point to open models becoming viable for token-efficient, long-running agentic engineering tasks. Source

2. NVIDIA Uses Vera CPU to Accelerate Chip Design Verification

NVIDIA. NVIDIA detailed how it is deploying the Vera CPU, built with 88 custom Olympus cores and an LPDDR5X memory subsystem, to speed up electronic design automation workflows for its next-generation processors. Working with Cadence and Synopsys, NVIDIA reported up to 1.5x higher performance on selected verification workloads such as Cadence Jasper formal verification and Synopsys VCS functional verification. The work underscores that simulation and verification remain CPU-bound stages where per-core performance and memory bandwidth drive turnaround time. Source

3. NVIDIA and Applied Materials Build a Digital Thread for Semiconductor Development

NVIDIA. NVIDIA and Applied Materials outlined an integrated, GPU-accelerated development model spanning atomic-scale materials discovery through full-fab optimization. The stack cites cuDSS delivering up to 10x speedups in sparse linear algebra, cuEST cutting density functional theory runs from five days on 64 CPU cores to about two hours on a single GPU, and PhysicsNeMo enabling up to 35x faster multiphysics chamber simulations, with Omniverse digital twins for factory layout. The approach ties Ginestra, CUDA-X libraries, PhysicsNeMo, and Omniverse into one continuous workflow for advanced chipmaking. Source